HAL Id: hal-01083038
https://hal-polytechnique.archives-ouvertes.fr/hal-01083038
Submitted on 14 Nov 2014
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Changes in rainfall regime over Burkina Faso under the climate change conditions simulated by 5 regional
climate models
Boubacar Ibrahim, Harouna Karambiri, Jan Polcher, Hamma Yacouba, Pierre Ribstein
To cite this version:
Boubacar Ibrahim, Harouna Karambiri, Jan Polcher, Hamma Yacouba, Pierre Ribstein. Changes in
rainfall regime over Burkina Faso under the climate change conditions simulated by 5 regional climate
models. Climate Dynamics, Springer Verlag, 2014, 42 (5-6), pp.1363-1381. �10.1007/s00382-013-1837-
2�. �hal-01083038�
Changes in rainfall regime over Burkina Faso under the climate change conditions simulated by 5 regional climate models
Boubacar Ibrahim•Harouna Karambiri• Jan Polcher•Hamma Yacouba• Pierre Ribstein
Received: 29 October 2012 / Accepted: 7 June 2013 / Published online: 20 June 2013 The Author(s) 2013. This article is published with open access at Springerlink.com
Abstract
Sahelian rainfall has recorded a high variability during the last century with a significant decrease (more than 20 %) in the annual rainfall amount since 1970. Using a linear regression model, the fluctuations of the annual rainfall from the observations over Burkina Faso during 1961–2009 period are described through the changes in the characteristics of the rainy season. The methodology is then applied to simulated rainfall data produced by five regional climate models under A1B scenario over two periods: 1971–2000 as reference period and 2021–2050 as projection period. As found with other climate models, the projected change in annual rainfall for West Africa is very uncertain. However, the present study shows that some features of the impact of climate change on rainfall regime in the region are robust. The number of the low rainfall events (0.1–5 mm/d) is projected to decrease by 3 % and
the number of strong rainfall events (
[50 mm/d) is expected to increase by 15 % on average. In addition, the rainy season onset is projected by all models to be delayed by one week on average and a consensus exists on the lengthening of the dry spells at about 20 %. Furthermore, the simulated relationship between changed annual rainfall amounts and the number of rain days or their intensity varies strongly from one model to another and some changes do not correspond to what is observed for the rainfall variability over the last 50 years.
Keywords
Climate change Regional climate model Rainy season Multiple linear regression Sahel Burkina Faso
1 Introduction
The first IPCC report on the climate change (Houghton
1990) has triggered a great interest in climate modeling inorder to understand climate mechanisms and to evaluate climate evolution at short and long terms under different climate change scenarios (Nakicenovic and Swart
2000;Solomon et al.
2007; Vanvyve et al. 2008). These simu-lations are implemented at different spatial scales, from the global to the regional, depending on the models and the aims of the studies. However, from regional to global simulations, all climate models project a warmer climate during the 21st century (Prabhakara et al.
2000; Wu et al.2007; Solomon et al.2007). Other climate parameters such
as rainfall are also projected to change from regional to global scale under a warming condition (Solomon et al.
2009; Wang et al.2009).
With a focus on West Africa, climate models project different trends for the annual rainfall amount over the 21st
B. Ibrahim (&)H. KarambiriH. YacoubaInstitut International d’Inge´nierie de l’Eau et de l’Environnement de Ouagadougou (2iE), BP 594, Ouagadougou 01, Burkina Faso
e-mail: boubacar.ibrahim@lmd.jussieu.fr Present Address:
B. Ibrahim
West African Science Service Center on Climate Change and Adapted Land Use (WASCAL), 06 BP: 9507, Ouagadougou, Burkina Faso
J. Polcher
Laboratoire de Me´te´orologie Dynamique du CNRS/IPSL, Universite´ P. & M. Curie (Paris 6), Tour 45, 3e`me e´tage, Case 99, 4 pl. Jussieu, 75252 Paris Cedex 05, France P. Ribstein
UMR 7619 Sisyphe (UPMC-CNRS-EPHE),
Universite´ P. & M. Curie (Paris 6), Case 123, 4 place Jussieu, 75252 Paris Cedex 05, France
DOI 10.1007/s00382-013-1837-2
century (Hulme et al.
2001; De Wit and Stankiewicz2006;Paeth et al.
2009). Table1summarizes the changes in the annual rainfall amount over the West African Sahel determined from several simulations. These studies show unfortunately a wide range of changes in the annual rainfall amount without any consensus either from the GCMs or from the RCMs.
Altogether, the climate models simulations do not show any consensus in the trends of the annual rainfall amount over West African Sahel during the 21st century even when they are run under the same climate change scenario at a high spatial resolution. With a more detail on the rainy season, a study performed with a regional climate model (REMO) under two scenarios, A1B (intermediate scenario) and B1 (low scenario), Paeth et al. (2009) found a weak change in precipitation over the middle of the current century and a lengthening of dry spells within the seasons.
This change of dry spells length within the rainy season despite the unchanged annual rainfall amount shows that an annual analysis of rainfall evolution can hide some changes in the internal of the rainy season that can have significant impacts on water availability and agricultural production.
Furthermore, Biasutti and Sobel (2009) found another change in the evolution of the characteristics of the rainy season from the CMIP3 rainfall. They found from an analysis of monthly data, a shortening of the rainy season over Sahel with a delayed season onset of the African monsoon during the 21st century. Hence, despite these disparities and the uncertainties of the climate models (d’Orgeval et al.
2006; De´que´ et al. 2007; Buser et al.2010) in the evolution of the annual rainfall amount for the
future period, a significant insight can be found on the characteristics of the rainy season. Thus, an investigation of the characteristics of the rainy season over the Sahelian region from a fine time step rainfall data is needed for a better understanding of the main changes in the rainy seasons over the future period.
On the other hand, from the observations, an analysis of the variability of rainfall regime over the region made by Le Barbe´ et al. (2002) and by Laux et al. (2009) showed that changes in two characteristics of the rainy season (number of rainfall events and the mean rainfall amount per event) over 1950–1990 provide an interesting results on this variability.
The decrease in annual rainfall amount over the region during the last four decades (Nicholson
2005; Lebel and Ali 2009; Mahe´ and Paturel2009) is characterized by a decreasein both rainfall frequency and intensity (Le Barbe´ et al.
2002;Balme et al.
2005) during the rainy season. However, therainfall frequency presents the most important contribution to the annual rainfall amount variability over Sahel. The impact of the rainfall frequency (number of rain days) on the annual rainfall amount variability was highlighted by an analysis of daily rainfalls over Niger (Le Barbe´ and Lebel
1997). Also, crops growth and hydrological cycle dependmore on rainy events organization in the rainy season than on the amount of the total rainfalls (Sivakumar
1992; Lebel andLe Barbe´
1997; Vischel and Lebel2007; Modarres 2010).Thus, an analysis of the evolution of rainfall regime over the Sahelian area from the characteristics of the rainy season better highlights the different changes in the rainfall pattern.
In this study, we analyze the evolution of rainfall regime over Burkina Faso, in West African Sahel, with regard to the changes in eight characteristics of the rainy season (date of the season onset, date of the end of season, season duration, number of rain days, mean daily rainfall, maximum daily rainfall, annual rainfall amount, and mean dry spell length).
These characteristics are determined throughout a discreti- zation procedure of the rainy season (Ibrahim et al.
2012).Indeed, the eight characteristics relate to the four main components of the rainy season: the rainy season period, the rainfall frequency and intensity and the dry spell lengths.
They describe overall the potentialities of the rainy season for crops growth and runoff processes (Barron et al.
2003;Balme et al.
2006), but for this analysis we consider seven Table 1 Changes in the annual rainfall amount over the 21st century from some studiesStudy Region Models Scenarios Change
Hulme et al. (2001) Central Sahel CCSR-NIES, CGCM1, CSIRO-Mk2, ECHAM4, GFDL-R15, HadCM2a, NCAR1
B1-low, B2-mid, A1-mid and A2-high
Significant increase
Cook and Vizy (2006) Sahel CM2.1, A2 Decrease
MIROC3.2, Significant increase
CGCM2.3.2 Slight decrease
Paeth and Hense (2004) Sahel ECHAM3 (coupled), ECHAM3/LSG and HADAM2
SST scenario and increase of the GHG
Slight decrease
Mariotti et al. (2011) West Sahel ECHAM5 and RegCM3 A1B Decrease
Diallo et al. (2012) West Sahel ICTP-RegCM3, MPI-REMO, METO HC-HadRM3P
A1B Significant decrease
SMHI-RCA Increase
These changes concern periods within 2030–2100 in comparison to 1970–2000 period
characteristics (the season duration is omitted because it comes from the date of the onset and the date of the end of season). Thus an assessment of the changes in these char- acteristics under the warmer conditions projected by the climate models will give a detailed insights into the overall impacts of climate changes on the rainy season. The changes in the seven characteristics of the rainy season under the climate change conditions for the IPCC A1B scenario over Burkina Faso are determined from rainfall data produced by five regional climate models (CCLM, HadRM3P, RACMO, RCA, and REMO) run over 1950–2050 period. Meanwhile, the changes in the rainy season are evaluated from a com- parison between the characteristics of the rainy season over the reference period of 1971–2000 and those over the pro- jection period of 2021–2050. For each period, a multiple linear regression model (Montgomery et al.
2001; Chen andMartin
2009) is used to describe the relationship between sixcharacteristics of the rainy season and the annual rainfall amount. The assessment of the different relationships would highlight the most important characteristics that significantly determine the evolution of the rainfall regime. So, the regression model method is first implemented on the observed data in order to verify whether the results presented by Le Barbe´ et al. (2002) for the Sahel are valid for the limited area of Burkina Faso or what has changed since 1990 (last year of Le Barbe´ et al. (2002) analysis).
2 Data and methodology
2.1 Data and study area
This study is based on daily rainfall amount recorded and simulated over Burkina Faso. The observed data over
1961–2009 period come from the national network of ten synoptic stations which are well spread over the country (Fig.
1). The advantage of this network is that all the stationspresent a daily time series of good quality over the consid- ered period (only one year, 1978 is missed at Bogande). The simulated data are produced by five RCMs run under the A1B scenario (Nakicenovic and Swart
2000) for 1950–2050period, except for RACMO which starts in 1970. Indeed, the five RCMs were chosen from a set of eleven RCMs which have been implemented by ENSEMBLE-Europe (http://
ensemblesrt3.dmi.dk/) for the AMMA program with a con-
dition that the simulations present full daily climate data (rainfall, temperature, humidity, wind speed and radiation) from 1971 to 2050. The list and detail references of the RCMs used are presented in Table
2. The RCMs simulationswere performed with a spatial resolution of 50 km
950 km and with two boundary conditions (Table
2). For CCLM,RACMO and REMO models, the large scale fields of the ECHAM5-r3 simulation are used as boundary conditions (Kjellstro¨m et al.
2011). But, the two other models, Had-RM3P and RCA, were run with the HadCM3Q0 outputs as boundary conditions (Kjellstro¨m et al.
2011). Furthermore,the five RCMs were implemented over a domain that covers the West African region, longitude: from 35 W to 31E and latitude: from 20S to 35N.
This analysis is based on the raw data from the five simulations in order to characterize the intrinsic changes that are projected by the RCMs under the A1B scenario over Burkina Faso and other similar climatic zone. Also, the characteristics of the rainy season are determined for each station from observed and simulated daily rainfalls.
However, our analyses over the whole country are based on the seasonal average value of the different characteristics over the ten stations.
Fig. 1 Synoptic stations and RCMsgrid boxesover Burkina Faso
2.2 Methodology
2.2.1 Definition of the characteristics of the rainy season and their variability
The rainy season over Burkina Faso is described throughout eight characteristics which highlight the main features and structure of the monsoon over the Sahelian area (Le Barbe´
and Lebel
1997; Sivakumar1988,1992; Barron et al.2003;Sultan and Janicot
2003; Ibrahim et al.2012): date of theseason onset (Onset), date of the end of season (End), season duration, number of rain days (NbRD), mean daily rainfall (MDR), maximum daily rainfall (MaxR), annual rainfall amount, and mean dry spell length (DryS). For this study, the dry spell is the number of consecutive days with rainfall amount lower than 0.1 mm (the lowest value recorded in the observations) both for the observations and the simulations.
The first two characteristics determined the period of the rainy season from which the other characteristics are derived.
However, several methods have been developed for the determination of the date of the season onset and the date of the end of season for the Sahelian area (Sivakumar
1992; Atiet al.
2002; Ibrahim et al. 2012). The statistic methoddeveloped by Ibrahim et al. (2012) is used in this study with a daily rainfall threshold of 0.1 mm/d. The criteria of this method are formulated as follows:
•
The season onset is determined after 5 % of the total annual rainfall amount is reached and the end of the season is determined after 95 % of the annual total rainfall amount has fallen;
•
The date of the season onset corresponds to the date of the rainfall higher than the average of annual first rainfall events over the entire period. In addition, to be considered, the rainfall event must not be followed by a dry spell longer than the median of the mean dry spell durations at the station or grid point;
•
The end of season is marked by a rainfall event occurring after or completing the 95 % of the annual rainfall amount and followed by a dry spell longer than the median dry spell duration at the station or grid point.
Meanwhile, the date of the season onset is a critical characteristic for the sowing period for food produc- tion while the second characteristic determines when the crops must reach their stage of maturity (Sivakumar
1992;Ati et al.
2002). Also, the rainy season period is delimitedby the date of the season onset and the date of the end of season (the two characteristics are determined from Ibra- him et al.
2012) from which the season duration is com-puted. Then, the following four characteristics describe the rainfall frequency and intensity which govern soil moisture and flow intensity along the rivers. Finally, the last char- acteristic, the mean dry spell length, quantifies the duration of the dry period between consecutive rainfall events.
Indeed, long dry spells in a rainy season can lead to crop drying out and poor harvests. Hence, characterizing the changes in these characteristics between two different periods may highlight the changes in the benefits of the rainy seasons in terms of available water resources and agronomic productions. Therefore, the significance of a change in each characteristic between the two periods is assessed with the Wilcoxon test of time series difference assessment (Ansari and Bradley
1960); for a given char-acteristic, the shift or difference between two periods is significant if the
pvalue is lower than 0.05.
Furthermore, the comparison periods for the observa- tions are determined through a statistical procedure which splits the full time series into periods of homogeneous data.
The procedure is called segmentation (Hubert et al.
1989).The segmentation procedure separates the observed annual rainfall amount time series into wet and dry periods with a significant difference in the magnitude of the annual rain- fall amounts for consecutive periods. The procedure is applied for the observed annual rainfall amount time series.
But, for the RCMs data, we consider two periods of comparison, the reference period of 1971–2000 and the projection period of 2021–2050. Indeed the projection period is taken with regard to its climate condition which is projected to be warmer than the reference period by the climate models under the climate change condition (Hulme et al.
2001; De Wit and Stankiewicz 2006; Paeth et al.2011).
2.2.2 Elaboration of the multiple linear regression of the annual rainfall amount
The annual rainfall amount is traditionally considered as the main characteristic of the rainy season (Ali and Lebel
2009; Lebel and Ali 2009; Mahe´ and Paturel 2009) fromwhich the variability of the rainfall regime is usually
Table 2 References of the fiveregional climate models Institute Model (RCM) Driving GCM Reference
HZG CCLM ECHAM5-r3 Rockel et al. (2008)
HC HadRM3P HadCM3Q0 Moufouma-Okia and Rowell (2009)
KNMI RACMO ECHAM5-r3 Meijgaard et al. (2008)
SMHI RCA HadCM3Q0 Samuelsson et al. (2011)
MPI REMO ECHAM5-r3 Kotlarski et al. (2010)
assessed. But, the annual rainfall amount is narrowly linked to the six characteristics of the rainy season (date of the season onset, date of the end of season, number of rain days, mean daily rainfall, maximum daily rainfall, and the mean dry spell length). In this study, the six characteristics are taken as predictors of the annual rainfall amount in order to get a more comprehensive understanding of the rainy season variability that cannot be achieved with the seasonal total rainfalls alone. Thus, a multiple linear regression procedure (Andrews
1974; Brown et al. 1998;Montgomery et al.
2001) is performed in order to repro-duce the annual rainfall amount from the six characteris- tics. This regression modeling aims to present a more complete picture of the rainy seasons over Burkina Faso from the observations during 1961–2009 period and from each of the five simulations. The multiple regression model is built from a sub-set of the six characteristics called the regression model’s pertinent variables. These pertinent variables have a none zero coefficients (Eq.
1) determinedfrom two methods over the target period: the deterministic method (Montgomery et al.
2001) and the Bayesian method(Chen and Martin
2009). However, it is presented here thedeterministic procedure based on the multiple linear regression model. The linear regression model of the annual rainfall amount is:
Pt¼fðXtÞ ¼CþX6
j¼1
ajxj:t ð1Þ
with
PðtÞthe annual rainfall amount for year
t(t the year index),
Xtvector of the regression model variables at year
t, j ð1j6Þ variable index, C constant of the regression model,
ajthe coefficient of variable
j, and xj.tvalue of variable
jfor year
t.So, the observed annual rainfall amount regression model is performed over the entire period of the observa- tions (1961–2009) in order to have a large sample, but for the RCMs, a regression model is calibrated over each period from the simulations (reference,
f1,
i=1 with t1
=1971–2000, and projection,
f2,
i=2 with t2
=2021–
2050) as it is assumed that the factors
ajand the coefficient C can vary between the two periods. The pertinent vari- ables of the regression model are determined through Stepwise procedure (Bendel and Afifi
1977) which elimi-nates variables that are not statistically significant in the model from the Akaike Information Criterion (this criterion gives the information lost for each candidate model and the pertinent model is the one with low AIC) (Seghouane and Amari
2007). In addition, the significance of the correlationbetween the selected pertinent variables is assessed with the Pearson test of correlation (Millot
2009). For this test,two variables are significantly correlated when the
pvalue is lower than 0.05 which is reached with a correlation
coefficient of about 0.6. So, the selected pertinent variables must be less correlated and represent the main character- istics of the rainy season that describe the structure of the rainy season over the given period. Also, the representa- tiveness of the regression model over the target period is assessed from its projections with the Wilcoxon test of the difference and the Pearson test for the interannual corre- lation. The linear regression model is considered valuable over a given period when there is no significant difference between its projections and the annual rainfall amount time series over the period. Then, the returned annual rainfall amount variance from the regression model, R-squared, is computed from the formula of Eq.
2(Scherrer
1984;Legendre and Legendre
1998). Altogether, the returnedvariance must be higher than 70 % (significant correlation with Pearson test) for a valuable regression model.
R2 ¼Xp
j¼1
bjqðP;xjÞ ð2Þ
with
bjthe standardized regression coefficient,
qðP;xjÞcorrelation coefficient between
P(annual rainfall amount time series over the considered period) and
xj(variable
jtime series over the considered period), the term
bj qðP;xjÞrepresents the contribution of variable
jto the restituted total variance
bjR2of the annual rainfall amount from the regression model.
On the other hand, as the regression models (f
1for the first period and
f2for the second period) are built from the time series of the six characteristics over each period and for each RCM, pertinent variables over the reference period and those over the projection period can be different for a given RCM.
Thus, the regression model of the reference period can be different with the regression model of the projection period with regard to the pertinent variables and the coefficients.
Over all, three cases of pertinent variables sets can be encountered in the two regression models establishment:
•
Same subsets of pertinent variables over the two periods, this implies no change in the main character- istics of the rainy season over the two periods (case 1);
•
The subset of the pertinent variables of one period is included in the subset of the pertinent variables of the other period; which mean that some changes in the rainy season structure may exist (case 2);
•
The two subsets of pertinent variables are different from one period to another, indicating a fundamental change in the structure of the rainy season (case 3).
Hence, these differences in the pertinent variables of the
two regression models highlight the change in the weight
of the relationship between the characteristics of the rainy
season and the annual rainfall amount. The significance of
the changes in the structure of the rainy season is assessed
from the performance of each regression model (f
1and
f2) over the two periods. This assessment helps also to select the most representative regression model over both refer- ence and projection periods with regard to the change in the annual rainfall amount. But, in case of the two regression models are not representative, a new regression model (f) is calibrated from the merging set of the pertinent variables over the two periods. So, if we call
X1the set of the per- tinent variables over the reference period and
X2the set of the pertinent variables over the projection period, the merged pertinent variables over the two periods are
X ¼X1UX2. The regression model
fis then elaborated from
X. Also, the significance of the contribution of the pertinent variables to the change in the mean annual rain- fall amount is assessed from a statistical analysis per- formed through the regression model
f.Indeed, for two different periods (significant change in the annual rainfall amount), the regression model (f) is applied from some substitutions of the data over the first period by the data over the second period. For each perti- nent variable j, its data over the first period are substituted by its randomly permuted data over the second period. The random permutation of the data is performed in order to break the interannual variability of the given variable over the second period. Thus, for each variable j, 1000 random permutations are performed in order to get a large sample of the data ranking. Then, for each variable j, the fictive projections (P
1;j) from
fare generated and compared to (P
2) (main projections over the period 2021–2050 from the regression model
f). Thus, variable j contributes signifi-cantly to the difference in the annual rainfall amount between the two periods if there is no significant difference between the fictive projections (P
1,j) and the main projec- tions (P
2). In addition, an assessment from some simulta- neous substitution of two or three variables is also done in case that no single substitution of the variables reproduces the change in the mean annual rainfall between the two periods. The simultaneous substitution of the variables consists in a substitution of the data of the considered variables at the same time with 1,000 random permutations of each variable.
2.2.3 Assessment of the contribution of the rainy season characteristics to the changes in the annual rainfall amount
The contribution of each variable to the mean deviation of the annual rainfall amount between the reference and the projection periods is assessed throughout the relative dif- ference of the annual rainfall amount
a¼ðPt2Pt1ÞPt1
.
Pt1mean annual rainfall amount over the first period and
Pt2mean annual rainfall amount over the second period. Let
a¼P6
j¼1
aj
with
ajthe contribution of variable
jto the rel- ative difference of the annual rainfall amount from Eq.
1.a¼ðPt2Pt1Þ Pt1 ¼X6
i¼1
ajðxj:t2xj:t1Þ Pt1 )aj¼ajðxj:t2xj:t1Þ
Pt1
ð3Þ
aj
coefficient of the variable
jin the regression model
f,xj:t1and
xj:t2average values of the variable over the first period and over the second period.
Also, the contributions of five daily rainfall classes to the relative change of the annual rainfall
d(Eq.
4) arecomputed from the basic data. The five rainfall classes (Ibrahim et al.
2012) considered in this study are: very low(0.1–5 mm/d), low (5–10 mm/d), moderate (10–20 mm/d), strong (20–50 mm/d) and very strong (
[50 mm/d). The contribution of each rainfall class to the change in the annual rainfall amount helps to identify which intensities determine the annual rainfall variability over Burkina Faso.
Indeed, the annual rainfall amount is computed from the five rainfall classes from
Pt¼P5k¼1
PCk:t
, with
PCk.tthe annual rainfall amount for the rainfall class
kand
Ptthe annual rainfall amount for year
t. Let Pt1be the average annual rainfall amount from the basic data over the period
i=1, 2 and
dckthe contribution of the rainfall class
k:d¼ðPt2Pt1Þ Pt1
¼X5
k¼1
ðPck:t2Pck:t1Þ Pt1
Or
d¼X5k¼1
dck)dck¼ðPck:t2Pck:t1Þ Pt1
ð4Þ
So,
dcdkrepresents the weight of the rainfall class
kto the whole variation of the annual rainfall amount between the two periods.
NB: All the analyses done in this study are performed with R software (http://www.r-project.org/).
3 Historical background of rainfall variability over Burkina Faso
In this section we focus our analysis on the characteristics
of the rainy season interannual variability in Burkina Faso
over the 1961–2009 period. The evolution of the rainfall
regime is characterized throughout the annual rainfall
amount variability in order to identify significant changes
that have occurred in the observed records and then their
relation with the six characteristics of the rainy season.
3.1 Annual rainfall variability over the period 1961–2009
An application of the segmentation procedure (Hubert et al.
1989) to the annual rainfall amount time series shows three
different homogeneous periods: 1961–1969, 1970–1990 and 1991–2009. The annual rainfall amount presents also dif- ferent coefficients of variation over the three periods: 8 % for the first period, 10 % for the second period and 11 % for the last period. So, the change in the coefficient of the inter- annual variability is not significantly different over the three periods. The annual rainfall amount mean decreases are 19 and 9 % respectively over the two last periods in comparison to the first period (Table
3). The three homogeneous periodsgiven by the segmentation procedure are in accordance with the results of Ali and Lebel (2009) who showed a rainfall decrease over the Sahelian area from the end of 1960s and the results of Nicholson (2005) and Mahe´ and Paturel (2009) who found that annual rainfall has increased over the Sahel since the end of 1990s. The three studies highlight that the 1970–1990 period was the driest period over Sahel during the last century. In addition, we compute the normalized index for annual rainfall, number of rain days and mean daily rainfall, with regard to the average over 1961–1969 period, in order to get the annual anomalies.
An analysis of the normalized indexes of three main characteristics of the rainy season (annual rainfall amount, number of rain days and mean daily rainfall) shows two main features for all variables (Fig.
2), a downward trend from1961 to 1984 and an upward trend from 1985 to 1995. We notice from Fig.
2, 3 years with an annual rainfall amountdeficit of more than 30 %: 1977, 1983 and 1984. The 1983 is the driest year during the second half of the twentieth century over Burkina Faso with a decrease of 34 % with regard to 1960s decade. Furthermore, it can be noted on this figure (Fig.
2) that the annual rainfall amount index is much bettercorrelated with the number of rain days index than with the mean daily rainfall index (correlation coefficient of 0.84 for the number of rain days against 0.47 for the mean daily rainfall). The correlation coefficient between the annual rainfall amount and the number of rain days is also higher than that between the annual rainfall amount and the mean daily rainfall over each of the three periods (higher than 0.6 for NbRD and lower than 0.6 for MDR). Hence, the analysis of the relationship between the annual rainfall amount and the six characteristics of the rainy season should lead for 1961–2009 period to a selection of the number of rain days as the dominant characteristic.
On the other hand, an assessment of the changes in the characteristics between two consecutive periods shows that only the end of season and the mean dry spell length have not significantly changed between the first and the second period (Table
3). But, for the changes between the secondand the third period, there is no significant change for four characteristics, the last two characteristics, the season onset and the mean daily rainfall. Altogether, in comparison with the first and the third period, the driest second period is characterized by a delayed season onset (short rainy sea- sons), a decrease in the number of rain days and in the intensity of the maximum daily rainfall (Table
3).From the Stepwise procedure of pertinent variables selection for a regression model, the overall six variables were selected to elaborate the regression model (Eq.
1)over the entire observation period (1961–2009). The cor- relation coefficients between the six variables are lower than 0.6, which means that the six variables are not closely linked to each other. The regression model reproduces 92 % of the observed annual rainfall variance with a partial contribution of the number of rain days of 56 %, 16 % for the mean daily rainfall, 11 % for the maximum daily rainfall and the other variables contribute at less than 10 %.
The multiple regression models’ projections present no significant difference with the observed annual rainfall amounts and present a correlation coefficient of about 0.9.
As verification, the Bayesian regression method (Chen and Martin
2009) was also used. It selects three pertinent variablesfor the regression model: the number of rain days, the mean daily rainfall and the maximum daily rainfall with a likelihood of 0.61 from a set of 10,000 iterations of the Markov Chain Monte Carlo (Gilks
1996). These three variables are alsofound to be dominant from the deterministic method, thus the Bayesian method confirms the relevance of these variables for the annual rainfall amount regression model. So, the multiple linear regression model built with the deterministic method is more suitable for the regression because of its simplicity and its appropriate description of the different changes in the evolution of the annual rainfall amount.
3.2 Description of the rainfall regime evolution during the period of 1961–2009
In order to compare the contribution of the changes in the various characteristics of the rainy season to the annual mean rainfall we will in the following use periods of equal length. First the pre-drought period (1961–1969) will be compared to the driest nine years during the drought (1977–1986). In a second step the recovery of rainfall will be examined with 19 years of the drought period (1972–1990) and the last segment of the time series with equal length (1991–2009).
3.2.1 Characterization of the annual rainfall amount decrease between 1961–1969 and 1977–1986
The analyzes of the change from the linear regression
model are performed between 1961–1969 period (P1) and
1977–1986 period (P2). Figure
3presents the annual rainfall amount projections from the regression model, the first two boxes (P1 and P2) are the projections from the master data of the two periods. The others boxes represent the projections from the pertinent variables with a substi- tution of the data (one, two and three variables) for the 1961–1969 period by the randomly permuted values over the 1977–1986 period. The magnitudes of the annual rainfall amount projected with the substituted pertinent variables (from Onset to DryS) are significantly higher than the annual rainfall amount of the dry period (P2). Overall, substituting NbRD produces the highest decrease in the annual rainfall amount, but not enough to reach the mag- nitude of the period 1977–1986 (Fig.
3). Thus, one variablecannot fully reproduce the decrease in the annual rainfall amount between the two periods. It was found that com- bining the substitution of NbRD with either MDR (NbMD) or MDR and MaxR (NbMDMa) was needed in order to reproduce the full magnitude of the rainfall reduction over P2. Even if the decrease in the annual rainfall amount is explained mainly by a decrease in number of rain days, the
magnitude of the decrease in the annual rainfall amount is obtained by the simultaneous impact of the decreases in three characteristics: number of rain days, mean daily rainfall and maximum daily rainfall.
The three variables contribute significantly to the annual rainfall amount decrease between the two periods. From Eq.
3, we compute a contribution of 58 % due to thenumber of rain days, 24 % due to the mean daily rainfall and 8 % due to maximum daily rainfall. So, the three variables reproduce about 90 % of the mean shift of the annual rainfall amount between the two periods. The sig- nificant delay of the season onset (Table
3) does not con-tribute significantly to the decrease in the annual rainfall amount because of the low correlation between the two characteristics (-0.25). Overall, from the regression model, the number of rain days represents the main char- acteristic that lowered the annual rainfall amount over the second period. This characteristic has decreased by about 15 % during the second period compared with the first period. The mean daily rainfall and the maximum daily rainfall have both decreased by 8 and 9 %, respectively.
Further to that, the annual rainfall amount decrease concerns all rainfall classes but at different levels. From Eq.
4, the contributions to the decrease in the annualrainfall amount are determined are presented in Table
4.So, the strong rainfall class determines the changes in the annual rainfall amount between the two periods with a contribution at about 57 %. However, the pattern is not the same for the decrease in the number of rain days, where the contributions of the four rainfall classes are significant.
Thus, the strongest changes in the number of rain days occur in the 10–50 mm/d part of the rainfall spectrum. In addition, a monthly analysis (not shown) of the number of rain days shows a significant decrease of rainfall frequen- cies at the core of the rainy season, June, July and August (decrease of about 15 % over the period 1977–1986 com- pared with the mean values over 1961–1969 period).
Table 3 Mean values of eight characteristics of the rainy season from the observations over the three main periods
1961–1969 1970–1990 1990–2009
Annual rainfall amount (mm/y) 895 722(219 %) 817(29 %)
Date of the season onset (days) 12/05 21/05 16/05
Date of the end of season (days) 30/09 28/09 30/09
Season duration (days) 141 130(28 %) 137(23 %)
Number of rain days (days) 53 44(217 %) 48(29 %)
Mean daily rainfall (mm/d) 14 13(27 %) 13 (-7 %)
Maximum daily rainfall (mm/d) 68 62(29 %) 68
Mean dry spell length (days) 3 3 3
The dates of the season onset and the end of season are represented as dates in the format dd/mm without year. The bold and italics values:
significant change from Wilcoxon test between consecutive periods. The values in bracket represent the relative variation in comparison with 1961–1969 period
Fig. 2 Evolution of the normalized indexes of three characteristics of the rainy season over Burkina Faso over the 1961–2009 period from the observations. Thevertical green linesrepresent the breaking years produced from the segmentation procedure, 1970 and 1990
3.2.2 Characterization of the annual rainfall amount increase between 1972–1990 and 1991–2009
Two periods of nineteen years are considered to analyze with the regression model the increase in rainfall from 1972–1990 (P1) to 1991–2009 (P2). Figure
4shows the impact of the different variables on the projected annual rainfall amount. In contrary to the previous analysis, here the annual rainfall amount projections with the NbRD reached the level of the median of the second period pro- jections even if their magnitude is not enough to reach the level of the third quartile (Fig.
4). The contributions of theother variables are significantly lower than for the second period annual rainfall (P2). Altogether, only projections from the combination of three variables (NbMDMa) reproduce the magnitude of the projections over the second period at a significant level (Fig.
4). For the changes inthese characteristics between the two periods, the number of rain days has increased by 8 % over the period 1991–2009. Also the maximum daily rainfall has increased by 9 % in contrary to the mean daily rainfall which still close to that over the 1972–1990 period. This comes from the discrepancies in the changes over the rainfall class mean intensities with a decrease in the low, the mean and
the strong rainfalls and an increase for the very strong rainfalls.
However, from the regression model, all the six perti- nent variables have contributed (Eq.
3) to the annualrainfall amount increase, with 8 % for the date of season onset, 2 % for the date of the end of the season, 60 % for the number of rain days, 6 % for the mean daily rainfall, 13 % for the maximum daily rainfall and 11 % for the mean dry spell length. So, as for the previous analysis on the description of the rainfall decrease, the number of rain days is the variable that contributes the most to the increase in the annual rainfall amount. Thus, even if the maximum daily rainfall has increased over the last period its impact on the annual rainfall amount remains lower than the impact of the number of rain days. On the other hand, the computation of the contribution of the five rainfall classes to the increase in annual rainfall (Eq.
4) shows that overallthe classes between 10 and 100 mm/d contributes about 90 % to the change in the annual rainfall amount (9 % due to the mean rainfall class, 58 % due to the strong rainfall class and 23 % due to the very strong rainfall). But, for the annual number of rain days, only the very low class has significantly increased by about 25 %. The other rainfall classes frequencies display small increases in their
Fig. 3 Impact of the characteristics of the rainy season on themagnitude of the annual rainfall amounts over the period 1977–1986 from the observations. The whisker boxes represent the full time series; the bottom whisker represents the minimum between the minimum of the time series and the median-1.5DQ (DQ represents the interquartile), the first quartile (25 %) is thebottomof thebox, the median (bold dash), the third quartile (75 %) is thetopof theboxand thetop whiskerrepresents the minimum between the maximum of the time series and the median?1.5DQ. P1 1961–1969 period, P2 1977–1986 period,Onsetsubstitution of the date of the rainy season onset,Endsubstitution of the date of the end of the rainy season,
NbRDprojections from the substitution of the number of rain days, MDR projections from the substitution of the mean daily rainfall, MaxRprojections from the substitution of the maximum daily rainfall, DrySprojections from the substitution of the mean dry spell length, NbMDprojections from the simultaneous substitution of the number of rain days and the mean daily rainfall,NbMDMaprojections from the simultaneous substitution of the number of rain days, the mean daily rainfall and maximum daily rainfall. Thered linesindicate the range of the annual rainfall amount of P2 to be reproduced by the regression model from the substitution of the pertinent variables. The bold boxindicates no significant difference with P2
Table 4 Contribution of the rainfall classes to the changes in the mean annual amount and the mean number of rain days between 1961–1969 and 1977–1986
Very low (%) Low (%) Moderate (%) Strong (%) Very strong (%)
Rainfall amount 3 7 18 57 15
Number of rain days 16 18 26 35 5
frequencies but remain lower than those over 1961–1969 period.
The results found in this analysis of daily rainfall cor- respond well to those Le Barbe´ et al. (2002) have obtained with another method (leak distribution model) and over the entire Sahel. This confirms the ability of the multiple linear regression model to describe in more detail the evolution of the mean annual rainfall amount and the different charac- teristics of the rainy season. This procedure will help us to better describe the different changes in the characteristics of the rainy season projected by the five regional climate models for a warmer climate.
4 Evolution of the rainfall regime from five RCMs
The climate evolution over the projection period is char- acterized from the change in the annual rainfall amount produced by the five RCMs. Figure
5presents the changes in mean annual rainfall produced by the five RCMs and also provides, as a reference, the observed values for the period 1971–2000. It can be noted that CCLM, HadRM3P, RACMO and REMO show a significant overestimation while RCA shows a significant underestimation of the annual rainfall amount for the reference period. Table
5summarizes the different deviations of the eight charac- teristics of the rainy season between the simulations and the observations over the reference period. A detailed analysis of these biases is presented in Ibrahim et al.
(2012). Meanwhile, the divergences between the five RCMs to reproduce the observations highlight the uncer- tainties in the climate models simulations to reproduce the West African Sahel rainfalls (Paeth et al.
2011; Karambiriet al.
2011).The whisker boxes of the annual rainfall amounts in Fig.
5show that changes in annual rainfall between the reference and the projection period are RCM dependent.
An assessment of the different changes highlights three possible cases: an increase for HadRM3P and RACMO, a decrease for CCLM and RCA, and no significant change for REMO. From this small sample it becomes evident that the influence of the driving GCM (Table
2) on the pro-jected rainfall changes by RCMs is small as for the first two cases different lateral boundary conditions (GCM
simulations) have been used (Jones et al.
1995, Mariottiet al.
2011). The increasing change concerns the twomodels which present the two highest mean annual rain- falls over the reference period and the decrease concerns the models which present the two lowest mean annual rainfalls over the reference period. However, the variance of the annual rainfall amounts is significantly homoge- neous over the two periods for each RCM with variance ratios between 0.8 and 1.3, and a
pvalue of Fligner-Killeen test (Fligner and Killeen
1976) higher than 10 %. So,despite the significant changes in the magnitude of the annual rainfall amounts, the variance of the annual rainfall amounts has not significantly changed between the two periods. The slight increases in the variances over the projection period for CCLM, HadRM3P and RACMO (Fig.
5) are not significant. These divergences between theRCMs in the evolution of the annual rainfall amount cor- respond to the results of previous studies for the West African region conducted with GCMs and RCMs (Hoerling et al.
2006; Paeth et al. 2009; Biasutti and Sobel 2009).However, despite the disagreement within the CMIP3 models in the evolution of the summer time total rainfall over the 21st century, Biasutti and Sobel (2009) found a robust delay of the rainy season onset in a warmer climate.
An analysis of the evolution of rainfall over Burkina Faso throughout the six main characteristics of the rainy season will better highlight the different changes in the rainfall regime even though the impacts on annual rainfall may be small or contradictory. The changes in the different char- acteristics of the rainy season will be evaluated with regard to the averages over the reference period presented in Table
5.4.1 Description of the evolution of annual rainfall amount predictors
4.1.1 Rainy season start and end dates
Changes in the dates of the season onset are model depen- dent. One model, CCLM shows a significant delay of about one week for the projection period while the other RCMs reveal no significant change (Table
6). HadRM3P, RACMOand REMO show a slight delay of few days (less than 4 days) on average while RCA shows no change in the mean date of
Fig. 4 Impact of thecharacteristics of the rainy season on the magnitude of the annual rainfall amounts over 1991–2009 period from the observations.P11972–1990 period,P21991–2009 period.
The other indications are the same as in Fig.3
the season onset. But for the end of the rainy season, the five RCMs show a general consensus of a delay which is signif- icant for the HadRM3P and RACMO models. The delay of the end of the rainy season in these two models is about one week. Changes in the dates of the end of the rainy season are not significant for CCLM, RCA and REMO; they present a slight delay of a few days on average (Table
6). As a con-sequence of the impact on the rainy season duration, only CCLM shows a significant shortening of the rainy season by one week, mainly due to the delayed onset. This is in agreement with the change in the observed season duration between the wet period 1961–1969 and the dry period 1970–1990. In contrast, the other RCMs present a slight extension of the rainy season by a few days (less than 4 days) on average. Indeed, from the Fligner-Killeen test, the vari- ances of the rainy season duration haven’t significantly changed between the two periods for the five models and the ratios of the variances are between 0.8 and 1.3. The signifi- cant delay of the end of the season observed for HadRM3P and RACMO is not enough to produce a significant length- ening of the rainy season because of the noise brought by the onset date. For these models, the rainy season period seems to be delayed without any change in the season duration.
Thus the rainy season period is not projected to change sig- nificantly in these two models despite their significant increase in the annual rainfall amount.
For the dry spell length evolution, a general consensus comes out of the five RCMs on a lengthening of the dry spells (Table
6). Two models, CCLM and RACMO CCLMand RACMO show a significant increase in the mean dry spell length of more than 5 %. The increase in the dry spell length has been found by Karambiri et al. (2011) from a different method of rainy season description. These chan- ges in the mean dry spell length have different origins;
decrease in number of rain days for CCLM while for RACMO the lengthening of the season duration is the likely cause. However, the mean dry spell length has remained stable over the observational record despite the significant changes in both rainy season duration and number of rain days.
4.1.2 Rainfall frequency and intensity
The number of rain days, the mean daily rainfall and the maximum daily rainfall allow us to better examine how the rain events change with climate. For the changes in the number of rain days (Table
6) only CCLM shows a sig-nificant decrease of around 14 % (a decrease of 6.4 days) which is in the range of the observed annual rainfall amount decrease over the two last periods (1970–1990 and 1991–2009) in comparison to the period 1961–1969. The other models only display small changes ranging from an
Fig. 5 Magnitudes of theannual rainfall amounts from the five RCMs over the reference and the projection periods. Same as Fig.3
Table 5 Mean values of the rainy season characteristics from the observations and the five RCMs over the reference period 1971–2000
OBS CCLM HadRM3P RACMO RCA REMO
Annual rainfall amount (mm/y) 760 840 1160 900 670 890
Date of the season onset (days) 138 152 105 129 121 143
Date of the end of season (days) 272 279 292 295 271 280
Season duration (days) 134 127 187 166 150 137
Number of rain days (days) 46 56 144 116 61 74
Mean daily rainfall (mm/d) 13 10.5 6.5 6 7 8
Maximum daily rainfall (mm/d) 64 113 68 92 70 138
Mean dry spell length (days) 3 2.5 2.5 2 2.5 2
increase of less than 1 % in HadRM3P and RACMO, and a decrease of about 3 % for RCA and REMO. In addition, we analyze the changes in the ratio of the number of rain days over rainy season length to describe how changes in the season duration impacts rainfall frequency. Altogether, there is no change in this ratio for HadRM3P and RACMO in contrast to a decrease in the ratio by 2 % for RCA and REMO and by 4 % for CCLM. Indeed, CCLM, the only model which has a significant shortening of the rainy season presents also the most important decrease in the proportion of rain days. For the mean daily rainfall (Table
6), only RACMO presents a significant change withan increase of 11 %. HadRM3P, CCLM and REMO pres- ent a slight increase of less than 6 % in contrast to RCA with a slight decrease of about 1 %. HadRM3P and RACMO display the same response with an increase in both number of rain days and mean daily rainfall while RCA shows a slight decrease in these two characteristics (Table
6). However, for CCLM and REMO a decrease inthe number of rain days can be observed while the mean daily rainfall increases leading to some compensation for the annual mean rainfall (Table
6). The five RCMs presentalso different changes in the evolution of the maximum daily rainfall. Only RACMO present a significant increase of about 30 %. Two models, HadRM3P and REMO present a slight increase (7 and 3 % respectively) while CCLM and RCA present a slight decrease by about 2 %. Thus, the changes in the three characteristics (number of rain days, mean daily rainfall and maximum daily rainfall) taken together are different from those observed because four models show an increase in the mean daily rainfall. Only CCLM presents a decrease in both number of rain days and maximum daily rain, consistent with the change found between 1961–1969 and 1970–1990 in the observational record (Table
5).On the other hand, the two models, CCLM and RCA, which present a decrease in the annual rainfall amount, present opposite signs in the evolution of the mean daily rainfall. In contrast, HadRM3P and RACMO with an increase in the annual rainfall amount present the same
type of change over all the seven characteristics (Table
6).Altogether, some consensuses are found on a delayed end of the seasons and a lengthening of the mean dry spells.
The first aspect has already been identified in GCMs (d’Orgeval et al.
2006; Biasutti and Sobel2009).4.2 Characterization of rainfall regime evolution from the simulations
The objective of this section is to estimate the contribution of each of the six characteristics of the rainy season to the change in annual mean rainfall determined in the five RCMs. The pertinent variables for the multiple linear regression model selected by the Stepwise procedure depend on the RCM and the target period (reference or projection). Table
7presents the contribution of each per- tinent variable to the total variance of annual rainfall (Eq.
3). The variances obtained from the regression models(f
1and
f2) are overall higher than 90 % which fulfills the requirement set for the methodology. The variance distri- bution of each regression model (Table
7) shows that themean daily rainfall is the most important variable for HadRM3P, RACMO, RCA and REMO models and account for more than 70 % of the total variance. But for CCLM, the mean daily rainfall dominates only during the reference period; the number of rain days becomes domi- nant during the projection period. Thus, the dominant variables in the regression model for annual rainfall of the five RCMs are not the same as for the observations (number of rain days is the dominant variable in that case).
Table
7shows also the modifications of the pertinent variables between the regression models for each RCM over the two periods. Only REMO presents the same per- tinent variables over the two periods. Even if there is a modification of the pertinent variables from one period to another for a given RCM, two variables, the number of rain days and the mean daily rainfall are always selected and account for more than 75 % of the total variance (Table
7).Thus, the number of rain days and the mean daily rainfall are the main variables in the projection of the annual
Table 6 Changes in the characteristics of the rainy season between 1971–2000 period and 2021–2050 period from the RCMs simulationsCCLM HadRM3P RACMO RCA REMO
Date of season onset (days) 18.6 ?0.5 ?0.5 -0.3 ?2.4
Date of end of season (days) ?3.5 18.4 16.1 ?5.9 ?6.8
Season duration (days) 27.9 ?3.6 ?3.0 ?3.7 ?0.8
Number of rain days (days) 27 ?0.4 ?1.3 -3.1 -2.1
Mean daily rainfall (mm/d) ?0.1 ?0.6 10.7 -0.1 ?0.3
Maximum daily rainfall (mm/d) -0.6 ?7.8 119.0 -1.8 ?2.1
Average length of dry spell (days) 10.2 ?0.1 10.2 ?0.1 ?0.1
The values represent the difference between the averages over the two periods. Bold and italics: significant change from Wilcoxon test
rainfall amount from the RCMs daily rainfall just as was found for the observations.
Furthermore, the performance of the two regression models (f
1and
f2) of each RCM is assessed over the two periods joined together (PP) in order to select the most valuable regression model. The selection criterion is based on the
pvalues of the test for the difference significance and the test for the correlation significance. So, the most valuable model is the one with a
pvalue close to 1 for the Wilcoxon test and a
pvalue close to 0 for the Pearson test.
But, in case none of the two regression models is valuable (significant difference and no significant correlation), a new regression model
fis generated from the union of the two periods (PP). Table
6presents the
pvalues of the two tests for the three regression models (f
1,
f2and
f). For fourRCMs: CCLM, HadRM3P, RACMO, and RCA, the regressions models
f1and
f2present significant differences with the regression model for the merged period (PP).
Hence, for these RCMs, the regression models are not valuable for the period they have not been calibrated on.
Thus, the modification of the pertinent variables in the regression models of these RCMs found in Table
8reveals a change in the structure of the rainy season.
In the same way as the analysis of the contribution of each variable to the overall variance done before (Table
9),here also the new regression models (f) reproduce over 94 % of the variance of the annual rainfall of the RCMs.
Table
9presents the contribution of each of the six char- acteristics (Eq.
3) to the relative difference of annualrainfall. This shows that the decrease in annual rainfall simulated by CCLM comes mainly from a decrease in the number of rain days and the delay of the season onset (Table
9). But, the decrease in the annual rainfall for RCAis explained by the decrease in both number of rain days and the mean daily rainfall amounts. In contrast, the increase of seasonal rainfall in HadRM3P and RACMO originate mainly in an increase of mean daily rainfall. The REMO model, which has no significant change in annual rainfall, is characterized by a positive contribution from the mean daily rainfall and a negative contribution from the number of rain days.
In the following analysis, performed with the regression model
f, the contribution of each of the six characteristics tothe change in annual mean rainfall is quantified through a random permutation of each characteristic for the projection period. Figure
6presents the annual rainfall amount pro- jected with different combinations of variables. For CCLM, the annual rainfall decrease is mostly explained by the impact of the change in the number of rain days which is the only variable that lowers significantly the annual rainfall from the level of P1 to the one of P2. For HadRM3P and RACMO, the increase in the annual rainfall amount over the second period can be attributed to changes in the mean daily rainfall. For RCA, the amplitude of the decrease in the annual
Table 7 Pertinent variables of the regression models and their contribution (in percentage) to the total variance of the annual rainfall amount
CCLM HadRM3P RACMO RCA REMO
P1 P2 P1 P2 P1 P2 P1 P2 P1 P2
Date of the season onset 18 x -0.2 -0.8 -7 8 x -1 -4 -9
Date of the end of season -0.8 x -0.2 -1.2 -0.1 1 x x 3 5
Number of rain days 35 52 21 10 4 8 9 8 18 7
Mean daily rainfall 42 35 76 89 93 75 89 86 73 75
Maximum daily rainfall x 5 x x 8 7 x 5 5 17
Mean dry spell length x x x 0.8 x -1 x x x x
Explained variance (%) 94 92 97 98 98 98 98 98 95 95
P1=Reference period 1971–2000, P2=Projection period 2021–2050, x=variable not selected for the regression model The bold values indicate the dominant variable in the regression model of the given period
Table 8 Performance of the regression models from thepvalues of Wilcoxon test and Pearson test
CCLM HadRM3P RACMO RCA REMO
f1 f2 f f1 f2 f f1 f2 f f1 f2 f f1 f2 f
Wp value Lv Lv 0.92 Lv Lv 0.74 Lv Lv 0.95 Lv Lv 0.73 0.3 0.3 0.94
Pp value Lv Lv \0.01 Lv Lv Lv Lv Lv Lv Lv Lv Lv Lv Lv Lv
Wp value=pvalue of Wilcoxon test, Pp value=pvalue of Pearson test,f1=regression model over the reference period,f2=regression model over the projection period,f=regression model over the two periods, Lv=\10-3
The bold values indicate the high returned variance over the three regression models
rainfall can only be reproduced by combining the number of the rain days and mean daily rainfall. Indeed, each of these variables has lowered the annual rainfall projected for the second period (Fig.
6). However, for REMO, the two vari-ables (the number of the rain days and the mean daily rain- fall) act on the annual rainfall amounts in opposite directions:
while the number of rain days lowers the mean annual rainfall, the mean daily rainfall increases it. Only the com- bination of the two variables produces the near zero change of total rainfall produced by this model. For all RCMs, the variance of the annual rainfall amount does not change sig- nificantly between the two periods as demonstrated by the Fligner-Killeen test.
Altogether, significant changes in annual rainfall amount produced by the five RCMs are dominated by changes in the number of rain days and/or the mean daily rainfall intensity. This corresponds to what has been found for the observation records as well and demonstrates that the models are able to pick-up this sensitivity of the rainy season of the Sahel.
4.3 Changes in daily rainfall for different intensities The changes in the total number of rain days and in the mean daily rainfall intensity are not homogeneously dis- tributed over the spectrum of the rainfall intensities. These changes may concern only part of the five rainfall classes defined in the methodology. Thus, for each RCM, the variation for each rainfall class is computed relative to the average over the five rainfall classes for the reference period. The relative variation for the rainfall class
kis:
kk¼DNcN k
1
100 with
DNckdifference of the numbers of rain days between the two periods for rainfall class
kand
N1average number of rain days over the reference period and a given RCM (Table
3). The same formula is used forthe mean daily rainfall, with
DPckthe difference of the mean daily rainfall and
P1the average of the mean daily for the reference period (Table
5).Table
10presents the relative variations in each rainfall class and RCM. CCLM presents the most important decrease in the number of rain days and this decrease
concerns all rainfall classes even if the very low rainfalls record the highest variation. But for the mean daily rainfall of this RCM, only the very strong rainfall class displays an increase. Thus, the slight increase in the mean daily rainfall presented by CCLM in Table
6is due to an increase in the intensity of rainfalls higher than 50 mm/d. The significant increase in the mean daily rainfall for HadRM3P and RACMO (Table
6) is mainly attributed to an increase inthe very strong rainfall intensity (Table
10). On the otherhand, Table
10shows that HadRM3P and RACMO, two models without any significant changes in the total number of rain days (Table
6), present two rainfall classes withsignificant change in the partial number of rain days.
Finally, for REMO, despite the decrease in the total num- ber of rain days, the very strong rainfall class has slightly increased in number in contrast to the other rainfall classes.
Altogether, two cases of one type of change are found with a decrease in the number of rain days over all rainfall classes for CCLM and RCA (Table
10). Thus, change inthe mean values of the number of rain days and in the mean daily rainfall does not mean a single type of change over all rainfall thresholds. Also, from Tables
5and
10two RCMs with the same type of change in the annual rainfall amount can present different combinations of type of change over the rainfall classes. However, the only consensus that comes out from the change in the five RCMs rainfall classes is a decrease in the number of the low rainfalls (0.1–5 mm/d). The second largest change concerns four RCMs (CCLM, HadRM3P, RACMO, REMO) with an increase in the very strong rainfalls (
[50 mm/d).
5 Summary
The structure of the rainy seasons is described in this study through a set of eight characteristics: date of the season onset (Onset), date of the end of season (End), season duration (SDR), number of rain days (NbRD), mean daily rainfall (MDR), maximum daily rainfall (MaxR), annual rainfall amount, and mean dry spell length (DryS). The seven char- acteristics address the main components of the rainy season
Table 9 Distribution of the changes in the annual rainfall amount within six characteristics of the rainy season for the five RCMsCCLM HadRM3P RACMO RCA REMO
Date of season onset (%) -2 -0.1 -0.3 ?0.04 -0.9
Date of the end of season (%) ?0.2 ?0.9 ?1.1 0 ?0.6
Number of rain days (%) 214 ?0.8 ?0.4 23.6 -2.3
Mean daily rainfall (%) ?0.3 15.2 18.8 -2.3 13
Maximum daily rainfall (%) 0 0 ?1.3 0 ?0.3
Mean dry spell length (%) 0 -0.1 -0.3 0 -0.2
The bold values represent the highest contribution for each RCM
over Sahel and allow to address properties of the rainy season more relevant for application as agricultural yields and water resources in the region. The characterization of the interan- nual variability of the observed and simulated rainfall over Burkina Faso is done with the multiple linear regression based on six characteristics of the rainy season (Onset, End, NbRD, MDR, MaxR, and DryS).
The linear multiple regression revealed that NbRD is the main characteristics of the rainy season that highlights the
different changes in annual rainfall amount over Burkina Faso during 1961–2009 period as was found in previous studies for the Sahelian area (Le Barbe´ et al.
2002). However,even if MDR has decreased during the drought period, it contributes less than NbRD to the variability of the annual rainfall amounts. Also, despite the significant increase in MaxR from the period 1970–1990 to the period 1991–2009 its contribution to the increase in the annual rainfall amount over the last two decades is less important than that from
Fig. 6 Impact of each variableon the change in the annual rainfall amount of each RCM between the reference and the projection periods. The point represents the average of the time series, thered line represents the level of the average of the projections of P2 from basic data and thewhiskers represent the standard deviation of the time series.P1Reference period 1971–2000,P2
Projection period 2021–2050